PCDE Optimising Service Performance Practice Test 4 — 15 Questions

Question 1

A DevOps team is deploying a web application on Google Kubernetes Engine (GKE) that experiences daily traffic spikes. They want to automatically adjust the number of pod replicas based on CPU utilization. Which Kubernetes resource should they use?

Accepted Answer

Horizontal Pod Autoscaler (HPA). The Horizontal Pod Autoscaler (HPA) automatically scales the number of pod replicas based on resource utilization metrics like CPU or memory. The Vertical Pod Autoscaler (VPA) adjusts resource requests/limits, not replica count. Cluster Autoscaler adjusts node count, not pods. PodDisruptionBudget controls voluntary disruptions.

Answer

Cluster Autoscaler

Answer

PodDisruptionBudget

Answer

Vertical Pod Autoscaler (VPA)

Question 2

A Cloud Run service handles HTTP requests that each involve a short background task after the response is sent. The service currently has CPU throttled when not handling requests, causing background tasks to fail. Which configuration change ensures background tasks complete?

Accepted Answer

Set CPU to 'always on'. Setting CPU to 'always on' prevents Cloud Run from throttling CPU when the container is not actively serving a request, allowing background tasks to complete. Min instances reduce cold starts but don't affect CPU throttling. Concurrency controls request handling. Execution environment gen2 offers higher memory but doesn't change CPU throttling behavior.

Answer

Set concurrency to 1

Answer

Set execution environment to gen2

Answer

Set min instances to 1

Question 3

A company stores time-series data in Cloud Spanner with a primary key composed of a timestamp prefix and a user ID suffix. They notice high write latency and hotspotting on a specific node. How should they redesign the primary key to distribute writes evenly?

Accepted Answer

Use a hash of the user ID as the prefix, then timestamp. Hotspots occur in Spanner when using monotonically increasing keys (like timestamps) because all writes go to a single tablet. Using a hash prefix or bit-reversal distributes writes across tablets. Keeping a timestamp suffix still causes hot leading key. None and using a secondary index doesn't fix the primary key issue.

Answer

Use a monotonically increasing counter as the prefix

Answer

Keep the current design but add a secondary index

Answer

Reverse the key order: timestamp suffix, user ID prefix

Question 4

A team runs a Cloud SQL MySQL instance serving an e-commerce application. Read traffic is growing and causing increased latency on the primary instance. They want to offload read queries with minimal application changes. Which solution should they implement?

Accepted Answer

Create Cloud SQL read replicas and direct read-only queries to them. Cloud SQL read replicas are designed to offload read traffic from the primary instance. The application can be configured to send read-only queries to the replica endpoint. Connection pooling with PgBouncer helps with connection management but does not offload reads. Query optimization reduces load but doesn't scale reads. Read replicas are the standard solution.

Answer

Upgrade the primary instance to a higher machine type

Answer

Optimize slow queries using EXPLAIN ANALYZE and add indexes

Answer

Enable connection pooling with Cloud SQL Auth Proxy and PgBouncer

Question 5

A GKE cluster runs a stateful workload that requires persistent volumes. The nodes are managed by a node pool with autoscaling enabled. During scale-down, the cluster autoscaler sometimes removes nodes that host critical pods with local data. How can the team prevent this?

Accepted Answer

Set a PodDisruptionBudget with maxUnavailable=0 for the critical workloads. PodDisruptionBudgets (PDB) allow specifying the minimum number of available pods during voluntary disruptions like cluster autoscaler scale-down. By setting a PDB with maxUnavailable=0, the autoscaler will not drain nodes that would violate the budget. Node taints and tolerations control scheduling but not disruption. Cluster autoscaler flags like scale-down-delay only delay scale-down, not prevent it for specific pods. VPA does not affect node selection.

Answer

Use node taints and tolerations to pin pods to specific nodes

Answer

Set the cluster autoscaler flag --scale-down-delay to a high value

Answer

Configure Vertical Pod Autoscaler to increase pod resources

Question 6

A company runs a batch processing job on Compute Engine VMs. The job is fault-tolerant and can handle individual VM failures by restarting tasks. To reduce costs, they want to use the cheapest possible VMs while ensuring the job completes within a flexible time window. Which VM option is MOST cost-effective?

Accepted Answer

Preemptible VMs. Preemptible VMs are significantly cheaper (up to 60-80% discount) than regular VMs but can be terminated at any time. Since the job is fault-tolerant, preemptibles are ideal for cost savings. Committed use discounts require a 1-year or 3-year commitment, less flexible. Standard VMs are more expensive. Sole-tenant nodes are for isolation, not cost savings.

Answer

Standard VMs with committed use discounts

Answer

Sole-tenant VMs

Answer

Standard VMs without discounts

Question 7

A Cloud Pub/Sub subscription is processing a high volume of messages, but the subscriber frequently receives messages that have already been processed (duplicates). The subscriber's processing time varies from 100 ms to 10 seconds. Which parameter should be adjusted to reduce duplicates while maintaining throughput?

Accepted Answer

Increase the acknowledgement deadline to 60 seconds. Pub/Sub redelivers messages if the acknowledgement deadline expires before the subscriber acknowledges. Setting a longer acknowledgement deadline (e.g., 60 seconds) gives the subscriber more time to process and acknowledge, reducing duplicates. Flow control limits outstanding messages but does not prevent duplicates. Using ordering keys ensures order but can increase duplicates due to head-of-line blocking. Increasing the number of pullers increases throughput but not directly reduce duplicates.

Answer

Enable flow control with a max outstanding message count

Answer

Use ordering keys on the subscription

Answer

Increase the number of puller clients

Question 8

A team is using Cloud Spanner for a global user database. They frequently run JOIN queries between a Users table and an Orders table. The queries are slow and the team suspects they are causing cross-node fan-out. Which schema design technique would reduce latency by co-locating related data?

Accepted Answer

Use interleaved tables (Orders interleaved in Users). Interleaved tables in Cloud Spanner store child rows physically with the parent row, allowing efficient JOINs without cross-node fan-out. Secondary indexes improve lookup but do not co-locate data. Hash-prefixed keys distribute writes but don't co-locate parent-child data. Denormalization avoids JOINs but may not be suitable for all cases.

Answer

Add a hash prefix to the primary key of both tables

Answer

Denormalize the Orders data into the Users table

Answer

Create a secondary index on the foreign key column

Question 9

A Cloud Run service experiences cold starts on the first request after being idle, causing latency spikes. The team wants to eliminate cold starts entirely. Which configuration setting should they use?

Accepted Answer

Set min instances to a value greater than 0. Setting a minimum number of instances (min instances) keeps the specified number of instances always warm, preventing cold starts. CPU always on prevents throttling but does not eliminate cold starts. Max instances limits scale but does not prevent cold starts. Concurrency controls how many requests an instance can handle but does not eliminate cold starts.

Answer

Set CPU to always on

Answer

Set max instances to 1

Answer

Set concurrency to 1

Question 10

A GKE cluster runs a mix of batch and latency-sensitive services. The batch jobs require occasional large CPU bursts, but the latency-sensitive services need consistent performance. The team wants to avoid CPU contention without over-provisioning nodes. Which approach should they take?

Accepted Answer

Create separate node pools: one with high-CPU machines for batch jobs and one with general-purpose machines for latency-sensitive services, and use taints/tolerations. Node pools with different machine types allow separating workloads onto appropriate hardware. Using taints and tolerations ensures that batch pods only run on the batch node pool, preventing interference with latency-sensitive services on the general pool. This approach avoids over-provisioning by right-sizing each pool. VPA and HPA adjust resources/replicas but do not prevent CPU contention between different workloads on the same node. PDBs control disruptions, not resource contention.

Answer

Configure a PodDisruptionBudget for the latency-sensitive services

Answer

Use Vertical Pod Autoscaler (VPA) for all pods

Answer

Use Horizontal Pod Autoscaler (HPA) with CPU target

Question 11

A company runs a Cloud SQL PostgreSQL instance for a SaaS application. They notice that the database CPU is consistently above 90% during peak hours, and queries slow down. The application is read-heavy and can tolerate some replication lag. Which action would MOST effectively reduce CPU load on the primary?

Accepted Answer

Create read replicas and route read-only queries to them. Creating one or more read replicas and offloading SELECT queries to them reduces CPU load on the primary instance. Adding more vCPU to the primary (scaling up) increases capacity but is more expensive and may not be as cost-effective as read replicas. PgBouncer helps with connection overhead but not CPU load from queries. Query optimization helps but may not reduce load enough if the volume is high.

Answer

Use EXPLAIN ANALYZE to optimize the slowest queries

Answer

Increase the number of vCPUs on the primary instance

Answer

Enable connection pooling with PgBouncer via Cloud SQL Auth Proxy

Question 12

A team is using Cloud Spanner with a primary key of UUID v4 values. They notice that read performance is suboptimal for range scans over a subset of keys. Which index strategy improves range scan performance?

Accepted Answer

Create a secondary index using INTERLEAVE IN the base table. A secondary index with an interleaved parent stores index entries in the same tablet as the base table rows, reducing cross-node fan-out for range scans. A hash prefix would help for point lookups but not range scans. Storing the index in a separate table (global index) increases latency. Dropping the index makes range scans worse.

Answer

Add a hash prefix to the primary key

Answer

Remove the UUID primary key and use a sequential key

Answer

Create a secondary index without interleaving (global index)

Question 13

A company runs a microservices application on GKE. They want to automatically adjust both the number of pods (for varying load) and the resource limits of individual pods (to avoid resource waste). Which two Kubernetes resources should they configure together? (Choose two.)

Accepted Answer

Vertical Pod Autoscaler (VPA). HPA adjusts the number of pod replicas based on metrics like CPU. VPA adjusts resource requests and limits to right-size pods. Together, they can complement each other: VPA provides recommended resource values, and HPA scales based on load. However, when using both, VPA should be in 'Off' mode for recommendations only, to avoid conflicts. Cluster Autoscaler adjusts nodes, not pods. PDB controls disruptions.

Answer

PodDisruptionBudget

Answer

Cluster Autoscaler

Answer

Node Auto-Provisioning

Question 14

A Cloud Pub/Sub subscription is used to ingest real-time events. The subscriber's processing rate is slower than the publish rate, causing messages to back up. The team needs to increase throughput without losing messages. Which three actions should they take? (Choose three.)

Accepted Answer

Increase the number of subscriber clients (parallel pull consumers). Increasing the acknowledgement deadline prevents premature redelivery but does not increase throughput. Using flow control (max outstanding messages) can actually reduce throughput. To increase throughput: increase the number of subscriber clients (more parallel pulls), use pull subscriptions (more efficient than push for high volume), and increase the max outstanding messages per client (allows more concurrent processing). Ordering keys are not needed for throughput and can reduce it.

Answer

Set ordering keys on the subscription

Answer

Use a push subscription instead of pull

Question 15

A company wants to optimize costs for their Compute Engine instances. They have a mix of workloads: some are fault-tolerant batch jobs, and others are stateful services requiring consistent uptime. Which two strategies should they use to reduce costs? (Choose two.)

Accepted Answer

Use committed use discounts for the stateful services. Preemptible VMs are cheaper and ideal for fault-tolerant batch jobs. Committed use discounts (1 or 3 years) offer significant savings for stateful services that run continuously. Active Assist recommendations help right-size instances. Using larger instances with fewer VMs may not always save costs. CPU always on is a Cloud Run setting, not relevant to Compute Engine cost optimization.

Answer

Enable node auto-provisioning in GKE

Answer

Replace all instances with larger machines to reduce total instance count

Answer

Use CPU always on for all instances